← @bgSponge

After Anthropic made Fable 5 permanently available in subscription plans, I noticed a large…

Summary

Lon Lundgren argues that Anthropic's Fable 5 performance dropped because users are served varying inference regimes rather than consistent model capabilities.

  1. Thinking token volume declined.

    Lundgren's five measurement methods showed August delivered dramatically fewer thinking tokens than July.

    August delivered dramatically fewer thinking tokens than July

  2. Reasoning performance fluctuates wildly.

    The model's ability to reason fell over the entire period and changed in multi-day episodes.

  3. Inference regimes vary by user.

    Even at xhigh or max effort levels, most model invocations received little to no thinking tokens.

  4. Benchmarks are no longer guaranteed.

    When the model does engage in longer thinking runs, it almost never reaches published benchmark levels.

  5. Access does not equal capability.

    Lundgren claims access to frontier models no longer guarantees the inference regime required to reproduce frontier capabilities.

    Access to frontier models no longer guarantees access to the inference regime

Read the original ↗Save to your Sponge

Summarized with Sponge. Quotes are the author’s own words.

More from @bg

All readings →